| Abstract: | Database and information technology has been widely used in process industries to manage data related to environmental performance. However, apart from being collected and archived for subsequent retrieval, the data has not been effectively exploited for improving environmental performance. In this paper we report our work on application of data mining and knowledge discovery technology to the analysis of a database of aqueous effluents from an organic manufacturing plant. The focus is on developing a software analyser for Microtox prediction. However, this methodology is applicable to any ecotoxicity measurement and will therefore offer a means of minimising difficult and tedious testing. Principal component analysis is used to pre-process the data for removing noise and reducing dimensionality. Automatic clustering is employed to group the multidimensional data into classes, and from each class training and testing data sets are selected for developing a back-propagation neural network to predict Microtox. The result shows that the software analyser is able to give satisfactory predictions for both training and test data. The errors for all the training and testing data are shown to satisfy a normal distribution. The software analyser is further used to carry out sensitivity studies in order to identify compounds responsible for observed toxicity value, based on which improved process operational strategies can be developed. Several approaches are investigated, including correlation coefficient analysis, sensitivity study based on differential analysis, one variable deletion, fuzzy curve approach and combination of the above with principal component analysis. |
